Intelligent fusion platform scheduling method and system based on artificial intelligence cluster

Abstract

According to the intelligent fusion platform scheduling method and system based on the artificial intelligence cluster, efficient cooperation of multiple AI clusters is achieved through a unified scheduling layer and a feature matching algorithm, the operation complexity is reduced, and the task processing precision and the resource utilization rate are improved. According to the scheme, task allocation is comprehensively managed through a core scheduling controller, and an optimal processing mode is dynamically selected according to a comparison result of a feature vector value of input data and a preset threshold value; and when the feature vector value does not reach a threshold value, sending the tasks to a plurality of AI clusters in parallel for processing by adopting a cyclic distribution mechanism, and updating the matching vector count according to a returned result until a condition is met. And meanwhile, the resource information used by each AI cluster is kept newest through an asynchronous information synchronization process. According to the technical scheme, the efficiency and accuracy of multi-AI cluster cooperative work are remarkably improved, and the problems of platform splitting, resource waste, data risk and the like in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence application technology, and in particular to an intelligent fusion platform scheduling method and system based on artificial intelligence clusters. Background Art

[0002] With the rapid development of artificial intelligence (AI) technologies, various AI platforms (such as natural language processing, image recognition, and conversation generation) have become key drivers of digital transformation across industries. Currently, a wide variety of AI platforms have emerged on the market, each with significant differences in functionality, operational methods, and data processing logic. While these platforms support intelligent needs in diverse fields, their fragmented and heterogeneous nature also presents numerous challenges.

[0003] In real-world applications, users often need to simultaneously utilize multiple AI platforms to meet diverse business needs. However, frequent platform switching not only increases operational complexity but can also lead to poor data flow and inefficiencies, and even cause data loss or processing errors due to platform compatibility issues. Furthermore, the lack of unified management for resource scheduling across different platforms makes efficient collaborative computing difficult.

[0004] Although there are some task scheduling and resource management solutions in the existing technology, these solutions are usually limited to a single platform or specific scenarios and lack the global scheduling capabilities for multi-platform and multi-model clusters.

[0005] Therefore, there is an urgent need for an intelligent fusion solution that can integrate multiple artificial intelligence platforms, achieve unified scheduling and efficient resource collaboration, so as to improve the overall efficiency and user experience of artificial intelligence clusters. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent fusion platform scheduling method and system based on artificial intelligence clusters, which solves the problems of low collaboration efficiency and uneven resource allocation of multiple AI platforms through unified scheduling and intelligent matching mechanisms, and realizes efficient task processing and resource optimization.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present application provides an intelligent fusion platform scheduling method based on an artificial intelligence cluster, comprising the following steps:

[0009] Step 1: Receive an AI service request submitted by the user through the client. The request includes the AI model type and input data selected by the user.

[0010] Step 2: Extract the feature type of the input data from the feature model library through the core scheduling controller and generate the corresponding feature vector value;

[0011] Step 3: Determine whether the eigenvector value is greater than a preset threshold;

[0012] If it is greater than the preset threshold, the request is sent to the target AI cluster that matches the feature type for processing;

[0013] If it is less than or equal to the preset threshold, execute the following loop process:

[0014] (a) Distributing the request to multiple AI clusters for parallel processing;

[0015] (b) Obtain the matching vector value returned by each AI cluster and update the matching vector count of the corresponding AI cluster type;

[0016] (c) recalculating the feature vector value based on the updated matching vector count, and again determining whether it is greater than the preset threshold;

[0017] When the recalculated feature vector value is greater than the preset threshold for the first time, the loop ends;

[0018] Step 4: Return the AI cluster processing results that ultimately meet the threshold conditions to the client.

[0019] In a preferred embodiment, the method for generating a feature vector value in step 2 includes: performing type calculation boundary algorithm processing using a large visual feature model and a large language feature model, generating a high-dimensional semantic vector through a BGE-M3 (BigGAN-Enhanced Multimodal Model 3) embedding model, and determining the type vector of the input data, specifically comprising the following steps:

[0020] Using a large visual feature model and a large language feature model to perform multimodal feature extraction on the input data, respectively, to obtain a visual feature vector and a language feature vector;

[0021] Inputting the visual feature vector and the language feature vector into the BGE-M3 embedding model, and generating a high-dimensional semantic vector through a cross-modal fusion algorithm;

[0022] Similarity matching is performed in a feature model library based on the high-dimensional semantic vector to determine the type vector of the input data.

[0023] In a preferred embodiment, the method for extracting the feature type of the input data in step 2 includes: constructing an inverted index structure based on the BM25 (Best Match 25) algorithm, calculating the type of the input data in the inverted index as a basis for determining the feature type, and specifically including the following steps:

[0024] Build an inverted index for predefined feature types (e.g., "natural language processing" and "image processing"), where each feature type is associated with a set of representative documents, and count the term frequency (TF) and inverse document frequency (IDF) of the representative documents;

[0025] Performing word segmentation processing on the input data to extract key feature terms;

[0026] Calculate the relevance score between the input data and documents of each feature type based on the BM25 algorithm;

[0027] The feature type with the highest correlation score is selected as the type attribution of the input data.

[0028] In a preferred embodiment, the default value of the preset threshold in step 3 is 0.7, and the preset threshold supports dynamic adjustment through a configuration interface.

[0029] In a preferred embodiment, the method for updating the matching vector count corresponding to the AI cluster type in step 3(b) is:

[0030] The matching vector value of the corresponding AI cluster type is increased by 1, and a new feature vector value is generated based on the weighted update results of all AI clusters.

[0031] In a preferred embodiment, the method further includes an asynchronous information synchronization sub-process, including:

[0032] Step A: Start the asynchronous process to synchronize data in the type resource library to ensure that the resource information used by each AI cluster is the latest version;

[0033] Step B: Distribute the synchronized resource information to all AI clusters;

[0034] Step C: Obtain the updated feature vector values from each AI cluster and cache them to optimize subsequent processing efficiency.

[0035] In a more preferred embodiment, the cache processing in step C includes: temporarily storing the feature vector value and directly calling the cached result in the subsequent process to reduce the time consumption of repeated calculations.

[0036] In a preferred embodiment, the multiple AI clusters distributed in step 3(a) include at least two of: a conversation interaction cluster based on ChatGPT technology, a natural language processing cluster, and an image processing cluster.

[0037] In a second aspect, the present application provides an intelligent fusion platform system based on an artificial intelligence cluster, comprising:

[0038] The presentation layer module includes the client application (APP), web interface (HTML) and third-party application interface, which is used to receive user requests and return processing results;

[0039] The business layer module includes a system information management module, an application module, a knowledge base module, and a plug-in tool module, wherein the knowledge base module stores structured knowledge and rule documents to support business decisions;

[0040] The unified scheduling layer module includes a core scheduling controller and a scheduling algorithm selector. The core scheduling controller is configured to perform the following operations:

[0041] Based on the comparison result of the feature vector value of the input data and the preset threshold, if the feature vector value is greater than the threshold, the request is sent to the target AI cluster that matches the feature type for processing;

[0042] If the eigenvector value is less than or equal to the threshold, the following loop process is executed:

[0043] a) Distribute requests to multiple AI clusters for parallel processing;

[0044] b) Obtain the matching vector value returned by each AI cluster and update the matching vector count for the corresponding AI cluster type;

[0045] c) recalculating the feature vector value based on the updated matching vector count and again determining whether it is greater than a preset threshold;

[0046] When the recalculated feature vector value is greater than the preset threshold for the first time, the loop ends;

[0047] The artificial intelligence cluster layer module includes the Dify module (multimodal processing unit), the ChartGPT module, and multiple extended AI cluster modules, which are used to perform natural language processing, dialogue generation, and diversified intelligent tasks respectively.

[0048] In a preferred embodiment, the core scheduling controller includes:

[0049] The feature extraction unit is used to construct an inverted index using the BM25 (BestMatch 25) algorithm, calculate the correlation score between the input data and the feature type, and call the BGE-M3 (BigGAN-EnhancedMultimodalModel3) embedding model for unmatched data to generate a high-dimensional (e.g., 1024-dimensional) semantic vector;

[0050] Vector calculation unit, which calculates feature vector values based on BM25 score or BGE-M3 semantic similarity;

[0051] The scheduling decision unit selects to allocate the request to a single AI cluster or multiple AI clusters according to the threshold determination result.

[0052] In a preferred embodiment, the scheduling algorithm selector is configured to dynamically select the optimal scheduling algorithm (such as shortest job first algorithm, weighted round-robin algorithm) based on task type (such as real-time requirements), current resource load status and historical processing efficiency data.

[0053] In a preferred embodiment, the system further includes an asynchronous information synchronization module, including:

[0054] The resource synchronization unit is used to synchronize the latest version of the type resource library to all AI clusters based on scheduled tasks or data update events;

[0055] A vector update unit, configured to collect the feature vector values returned by each AI cluster and update the matching weights of the feature model library based on the feature vector values;

[0056] The cache computing unit uses distributed storage technology to temporarily store the updated feature vector values and directly calls the cached results in subsequent request processing to reduce computing time.

[0057] In a more preferred embodiment, the cache computing unit of the asynchronous information synchronization module adopts distributed storage technology (such as Redis) to achieve fast reading and writing and redundant backup of feature vector values.

[0058] In a preferred embodiment, the Dify module of the artificial intelligence cluster layer module supports multimodal data processing, including joint analysis of text, image and speech.

[0059] In a third aspect, the present application provides a computer storage medium, which is used to store computer execution instructions; the computer execution instructions are used to execute an intelligent fusion platform scheduling method based on artificial intelligence cluster provided in the first aspect of the present application.

[0060] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0061] The present invention provides a scheduling method and system for an intelligent fusion platform based on artificial intelligence clusters. Through a unified scheduling layer and feature matching algorithm, efficient collaboration of multiple AI clusters is achieved, operational complexity is reduced, and task processing accuracy and resource utilization are improved. This solution manages task allocation through a core scheduling controller, and dynamically selects the optimal processing method based on the comparison result of the feature vector value of the input data and the preset threshold. When the feature vector value does not reach the threshold, a circular distribution mechanism is adopted to send the task to multiple AI clusters for processing in parallel, and the matching vector count is updated according to the return result until the conditions are met. At the same time, an asynchronous information synchronization process is used to ensure that the resource information used by each AI cluster is kept up to date. The technical solution of the present invention significantly improves the efficiency and accuracy of the collaborative work of multiple AI clusters, and solves the problems of platform fragmentation, resource waste and data risks in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 Schematic diagram of the system architecture of a preferred embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the artificial intelligence cluster takeover process of a preferred embodiment of the present invention.

[0065] Figure 3 It is a schematic diagram of the working process of the core scheduling controller of the preferred embodiment of the present invention.

[0066] Figure 4 This is a flowchart of asynchronous information synchronization according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the above and other features and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for the purpose of explaining to those skilled in the art and are only exemplary and not restrictive.

[0068] It should be noted that the terms "first," "second," and the like in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this manner are interchangeable under appropriate circumstances. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0069] Example 1

[0070] This embodiment describes in detail the implementation process of the intelligent fusion platform scheduling method based on artificial intelligence cluster. Figure 1 The system architecture diagram shown and Figure 2 The AI cluster takeover process diagram shown in the figure below implements efficient collaboration among multiple AI clusters through a unified scheduling layer and intelligent matching mechanism. The method specifically includes the following steps:

[0071] Step 1: Request reception.

[0072] Users submit AI service requests through a client application (APP), web interface (HTML), or third-party application interface. The request includes the AI model type selected by the user (such as "image classification" or "natural language generation") and the input data to be processed (supporting multi-modal data upload, such as text, images, and voice).

[0073] For example, in a medical imaging diagnosis scenario, a doctor uploads a CT scan image through a web interface and selects the "lung disease classification" service.

[0074] Step 2: Feature extraction and vector generation.

[0075] After receiving the request, the core scheduling controller performs the following operations (see Figure 2-Figure 3 ):

[0076] (1) Feature type extraction based on BM25 algorithm

[0077] The BM25 algorithm (Best Match 25) is a statistical relevance scoring algorithm used to evaluate the relevance between documents and query terms.

[0078] The BM25 algorithm is used to construct an inverted index, analyze the metadata and keywords of the input data, and calculate its type in the feature model library. Specifically, the following steps are performed:

[0079] Inverted index construction involves building an inverted index for predefined feature types (e.g., "natural language processing" or "image processing"). Each feature type is associated with a set of representative documents (e.g., medical imaging reports or text corpora). The term frequency (TF) and inverse document frequency (IDF) of the terms in the documents are counted, and the term weights are calculated.

[0080] Feature type determination involves performing word segmentation on the input data (e.g., metadata from CT images) and extracting keywords (e.g., "pulmonary nodules," "high-density shadows"). The relevance score between the input data and documents of each feature type is calculated using the BM25 formula. If the highest score exceeds a preset threshold (default 0.7), it is directly assigned to the corresponding AI cluster. If the threshold is not reached, the system enters the BGE-M3 processing stage.

[0081] (2) Feature vector generation based on the BGE-M3 model

[0082] The BGE-M3 model (BigGAN-Enhanced Multimodal Model 3) is a multimodal embedding model that supports cross-modal fusion of visual and language features.

[0083] Use the BGE-M3 embedding model to map the input data into a high-dimensional semantic vector space and generate feature vector values, including:

[0084] Multimodal feature extraction: Visual feature vectors (such as shape and texture) are extracted from input data (such as images) through a large visual feature model. Language feature vectors (such as keywords and semantic relationships) are extracted from associated text (such as image descriptions) through a large language feature model.

[0085] Cross-modal fusion: The visual and language feature vectors are input into the BGE-M3 model, and a high-dimensional semantic vector is generated through the fusion algorithm to represent the comprehensive characteristics of the input data.

[0086] Step 3: Threshold determination and task distribution.

[0087] The core scheduling controller compares the generated feature vector value with the preset threshold (default 0.7, dynamically configurable):

[0088] If the eigenvector value exceeds the preset threshold, the request is sent directly to the matching AI cluster for processing. For example, if the eigenvector value of a lung CT image reaches 0.75, it is directly assigned to the "pneumonia diagnosis cluster."

[0089] If the eigenvector value is less than or equal to the threshold, the loop processing flow is started:

[0090] (a) Distribute requests to multiple AI clusters (AI cluster A, AI cluster B, ..., AI cluster N) in parallel;

[0091] (b) Receive the matching vector value returned by each cluster (e.g., the matching value of the pneumonia diagnosis cluster is 0.65), and update the matching count of the corresponding cluster (matching value + 1);

[0092] (c) Generate a new feature vector based on the updated matching value weights and determine again whether it meets the threshold.

[0093] Loop termination condition: When the recalculated eigenvector value exceeds the preset threshold for the first time, the loop ends.

[0094] Step 4: The result is returned.

[0095] The AI cluster processing results that ultimately meet the threshold conditions are returned to the client.

[0096] For example, when the loop process determines that the "pneumonia diagnosis cluster" is the best choice, its diagnosis report is returned to the doctor's workstation.

[0097] Step 5: Asynchronous information synchronization.

[0098] like Figure 4 As shown in the flowchart, the system regularly executes the following asynchronous processes to achieve dynamic updates of the type resource library and data collaboration between AI clusters, ensuring resource consistency and processing efficiency. The specific process is as follows:

[0099] (1) Type resource library synchronization:

[0100] The system synchronizes the resource library periodically or based on events to ensure that the classification standards, model parameters, and rule documents used by all AI clusters are the latest versions.

[0101] (2) Resource information is distributed to the AI cluster:

[0102] The synchronized resource information is distributed to different AI clusters for processing, and different clusters can perform different computing or analysis tasks.

[0103] (3) Update of eigenvector value:

[0104] Receive the matching vector value returned by each cluster (such as the matching value of the pneumonia diagnosis cluster is 0.65), and update the matching count of the corresponding cluster (matching value + 1) (such as the matching count of the "pneumonia diagnosis" cluster increases from 100 times to 101 times).

[0105] (4) Cache vector calculation ( Figure 4 The dotted part in the figure indicates an optional or auxiliary process):

[0106] To reduce the time spent on repeated calculations, the system can cache vector values so that the cached results can be directly called in subsequent requests, optimizing calculation efficiency.

[0107] In the above description, the feature vector value is the numerical representation of input data (such as text or images) converted by algorithms (such as the BM25 or BGE-M3 models) to quantify the core attributes of the data. The matching vector value is the confidence score returned by the AI cluster after processing the data, reflecting the cluster's suitability for handling the current task. The system determines the task distribution method by comparing the feature vector value with a preset threshold and selects the optimal cluster based on the matching vector value. The two work together to achieve precise scheduling.

[0108] According to an embodiment of the present invention, a computer storage medium is further provided for storing computer-executable instructions for executing an intelligent fusion platform scheduling method based on artificial intelligence clusters provided by the present invention.

[0109] Example 2

[0110] This embodiment describes in detail the specific architecture of the intelligent fusion platform system that implements the above method. Figure 1 As shown, the system adopts a layered design, and each module works together:

[0111] 1. Presentation layer module

[0112] Client application (APP): Provides a visual interactive interface that supports users to submit requests and view results.

[0113] Web interface (HTML): A web-based interactive interface suitable for access by various browsers.

[0114] Third-party application interface: connect to external third-party platforms or systems to achieve function expansion, data interoperability, and integrate multi-party service capabilities.

[0115] 2. Business layer module

[0116] System information management module: maintains platform configuration parameters (such as threshold settings), user permissions and other basic data.

[0117] Application module: encapsulates specific business logic and implements business functions (such as data processing, process control, etc.). It is the core implementation unit of business needs.

[0118] Knowledge base module: stores structured knowledge (such as medical diagnostic standards), business rule documents, etc., provides knowledge support for the system, and assists business decision-making or intelligent interaction.

[0119] Plug-ins and tool modules: Integrate various functional components (such as tools and extension plug-ins) in a pluggable manner, and support dynamic loading and unloading.

[0120] 3. Unified scheduling layer module

[0121] Core scheduling controller, including:

[0122] (1) Feature extraction unit: used to construct an inverted index using the BM25 algorithm, calculate the correlation score between the input data and the feature type, and call the BGE-M3 embedding model for unmatched data to generate a 1024-dimensional high-dimensional semantic vector.

[0123] (2) Vector calculation unit: Calculates feature vector values based on BM25 score or BGE-M3 semantic similarity.

[0124] (3) Scheduling decision unit: Based on the threshold judgment result, it selects to allocate the request to a single target AI cluster or multiple AI clusters.

[0125] Scheduling algorithm selector: Dynamically selects the optimal scheduling algorithm based on real-time system load and task characteristics (such as the shortest job first algorithm to prioritize lightweight tasks and the polling algorithm to balance cluster load), optimizing resource allocation efficiency and task processing performance.

[0126] 4. Artificial Intelligence Cluster Layer Module

[0127] Dify module: An AI capability execution unit, which can be a specific AI model or framework used to process natural language, image, and other intelligent tasks.

[0128] ChartGPT module: An artificial intelligence module based on ChatGPT technology, focusing on natural language processing scenarios such as dialogue interaction and text generation.

[0129] AI Cluster N: refers to other AI models or clusters, supporting the expansion of more types of AI capabilities to meet diverse intelligent processing needs.

[0130] Example 3

[0131] This embodiment further illustrates the execution process of the technical solution of the present application with reference to a specific example.

[0132] Scenario Background: A hospital uses an intelligent fusion platform to handle routine medical imaging diagnostics, requiring CT image classification (e.g., lung disease identification). Traditionally, doctors have to manually switch between multiple independent systems (e.g., image preprocessing systems and specialized diagnostic systems), resulting in low efficiency.

[0133] The execution process of the technical solution of this application is described in combination with this application scenario:

[0134] Step 1: User submits a request.

[0135] Operation: The doctor uploads a chest CT image through the platform's web interface (display layer module) and selects the "Chest Image Differential Diagnosis" service.

[0136] Input data: CT image files in DICOM format, including standard metadata (such as scanning parameters and patient information).

[0137] Step 2: Feature extraction and vector generation.

[0138] The core dispatch controller performs the following processing:

[0139] 1) Feature type extraction:

[0140] The BM25 algorithm was used to construct an inverted index, analyze the image metadata (such as scanning parameters and basic patient information), and extract key feature terms (such as "ground-glass opacity of the lungs" and "air bronchial sign").

[0141] Calculate the BM25 correlation score between the input data and predefined feature types (such as "lung infection" and "interstitial pneumonia"). If the highest score is greater than 0.7, it is directly assigned to the corresponding cluster; in this case, the score is 0.65, and it enters BGE-M3 processing.

[0142] 2) Generate feature vectors:

[0143] The texture features of CT images (such as ground-glass shadow distribution and density) are analyzed through large visual models (such as ResNet-50).

[0144] The semantic information of the associated text is parsed through a large language model (such as BERT).

[0145] The visual and text features are input into the BGE-M3 model to generate a high-dimensional semantic vector with an initial feature vector value of 0.65 (the preset threshold is 0.7).

[0146] Step 3: Threshold determination and task distribution.

[0147] The system starts the intelligent scheduling process:

[0148] 1) Parallel distribution:

[0149] Because the initial feature vector value (0.65) is less than the preset threshold (0.7), the request is distributed by the core scheduling controller to three AI clusters for parallel processing:

[0150] Dify multimodal clustering: Analyzing image texture features and correlating diagnostic reports.

[0151] Cluster A (general image classification cluster): general image classification model based on ResNet-50.

[0152] Cluster B (Chest Specialty Diagnosis Cluster): A deep learning model optimized for chest diseases.

[0153] 2) Processing and returning results:

[0154] The Dify cluster returned an eigenvalue of 0.68 (identifying “interstitial changes in the lungs”);

[0155] Cluster A returned an eigenvalue of 0.63 (classified as “common inflammatory changes”);

[0156] Cluster B returns an eigenvalue of 0.72 (matching "lung infectious lesions").

[0157] 3) Update count:

[0158] The core scheduling controller collects the matching vector values returned by each cluster, adds 1 to the matching vector count of cluster B, and recalculates the feature vector value to 0.73.

[0159] Step 4: The loop terminates and the result is returned.

[0160] Re-judge the threshold: the updated feature vector value (0.73) > the preset threshold (0.7), and terminate the loop.

[0161] Establish directional routing: Subsequent similar image requests are sent directly to cluster B (chest specialist diagnosis cluster), skipping the parallel distribution step.

[0162] Diagnostic report generation: The platform returns a structured report to the doctor client, including the diagnosis conclusion ("There is a high possibility of infectious lung lesions") and recommendations ("Recheck after anti-infection treatment is recommended").

[0163] Step 5: Asynchronous information synchronization.

[0164] 1) Knowledge base synchronization: The platform synchronizes the latest medical diagnostic standards regularly (e.g., every 10 minutes) and dynamically updates the feature model library of each AI cluster.

[0165] 2) Cache optimization: The characteristic value of cluster B (0.72) is cached in Redis distributed storage. The next similar request can be directly called, greatly shortening the response time.

[0166] This technical solution achieves efficient collaboration among multiple AI clusters through a unified scheduling layer and intelligent routing mechanism: doctors only need to submit images through a unified interface to complete the entire diagnosis process, completely solving the tedious problem of switching between multiple platforms; the system adopts a dynamic resource allocation strategy, calling multiple professional clusters in parallel for collaborative analysis at the initial low confidence level, and automatically routing to the optimal specialist cluster when the feature matching degree reaches the threshold, which not only ensures diagnostic efficiency but also avoids the risk of misjudgment of general models; through an asynchronous synchronization mechanism, it ensures that all diagnostic clusters obtain the latest medical standards and feature labels in real time, effectively preventing diagnostic bias caused by inconsistent data versions.

[0167] While the specific embodiments of the present invention have been described in detail above, these are merely exemplary and the present invention is not limited thereto. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, any equivalent changes and modifications made without departing from the spirit and scope of the present invention are intended to be encompassed within the scope of the present invention.

Claims

1. An intelligent fusion platform scheduling method based on artificial intelligence cluster, characterized in that: The steps include: Step 1: Receive an AI service request submitted by the user through the client. The request includes the AI model type and input data selected by the user. Step 2: Extract the feature type of the input data from the feature model library through the core scheduling controller and generate the corresponding feature vector value; Step 3: Determine whether the eigenvector value is greater than a preset threshold; If it is greater than the preset threshold, the request is sent to the target AI cluster that matches the feature type for processing; If it is less than or equal to the preset threshold, execute the following loop process: (a) Distributing the request to multiple AI clusters for parallel processing; (b) Obtain the matching vector value returned by each AI cluster and update the matching vector count of the corresponding AI cluster type; (c) recalculating the feature vector value based on the updated matching vector count, and again determining whether it is greater than the preset threshold; When the recalculated feature vector value is greater than the preset threshold for the first time, the loop ends; Step 4: Return the AI cluster processing results that ultimately meet the threshold conditions to the client.

2. The method according to claim 1, characterized in that The method for generating the feature vector value in step 2 includes: using a large visual feature model and a large language feature model to perform a type calculation boundary algorithm, generating a high-dimensional semantic vector through a BGE-M3 embedding model, and determining the type vector of the input data, specifically including the following steps: Using a large visual feature model and a large language feature model to perform multimodal feature extraction on the input data, respectively, to obtain a visual feature vector and a language feature vector; Inputting the visual feature vector and the language feature vector into the BGE-M3 embedding model, and generating a high-dimensional semantic vector through a cross-modal fusion algorithm; Similarity matching is performed in a feature model library based on the high-dimensional semantic vector to determine the type vector of the input data.

3. The method according to claim 1, characterized in that The method for extracting the feature type of the input data in step 2 includes: constructing an inverted index structure based on the BM25 algorithm, calculating the type of the input data in the inverted index as a basis for determining the feature type, and specifically including the following steps: Build an inverted index for predefined feature types, where each feature type is associated with a set of representative documents, and count the frequency and inverse document frequency of terms in the representative documents; Performing word segmentation processing on the input data to extract key feature terms; Calculate the relevance score between the input data and documents of each feature type based on the BM25 algorithm; The feature type with the highest correlation score is selected as the type attribution of the input data.

4. The method according to claim 1, wherein The method for updating the matching vector count corresponding to the AI cluster type in step 3(b) is: The matching vector value of the corresponding AI cluster type is increased by 1, and a new feature vector value is generated based on the weighted update results of all AI clusters.

5. The method according to claim 1, wherein The method further includes an asynchronous information synchronization sub-process, including: Step A: Start an asynchronous process to synchronize data in the type resource library to ensure that the resource information used by each AI cluster is the latest version; Step B: Distribute the synchronized resource information to all AI clusters; Step C: Obtain updated feature vector values from each AI cluster and cache them to optimize subsequent processing efficiency; The cache processing in step C includes temporarily storing the feature vector values and directly calling the cached results in subsequent processes to reduce the time consumption of repeated calculations.

6. An intelligent fusion platform system based on artificial intelligence cluster, characterized in that: include: The presentation layer module includes client applications, web interfaces, and third-party application interfaces, and is used to receive user requests and return processing results. The business layer module includes a system information management module, an application module, a knowledge base module, and a plug-in tool module, wherein the knowledge base module stores structured knowledge and rule documents to support business decisions; The unified scheduling layer module includes a core scheduling controller and a scheduling algorithm selector. The core scheduling controller is configured to perform the following operations: Based on the comparison result of the feature vector value of the input data and the preset threshold, if the feature vector value is greater than the threshold, the request is sent to the target AI cluster that matches the feature type for processing; If the eigenvector value is less than or equal to the threshold, the following loop process is executed: a) Distribute requests to multiple AI clusters for parallel processing; b) Obtain the matching vector value returned by each AI cluster and update the matching vector count for the corresponding AI cluster type; c) recalculating the feature vector value based on the updated matching vector count and again determining whether it is greater than a preset threshold; When the recalculated feature vector value is greater than the preset threshold for the first time, the loop ends; The artificial intelligence cluster layer module includes the Dify module, the ChartGPT module, and multiple extended AI cluster modules, which are used to perform natural language processing, dialogue generation, and diversified intelligent tasks respectively.

7. The system according to claim 6, characterized in that The core scheduling controller includes: The feature extraction unit is used to construct an inverted index using the BM25 algorithm, calculate the correlation score between the input data and the feature type, and call the BGE-M3 embedding model to generate high-dimensional semantic vectors for unmatched data; Vector calculation unit, which calculates feature vector values based on BM25 score or BGE-M3 semantic similarity; The scheduling decision unit selects to allocate the request to a single AI cluster or multiple AI clusters according to the threshold determination result.

8. The system according to claim 6, characterized in that The system also includes an asynchronous information synchronization module, including: The resource synchronization unit is used to synchronize the latest version of the type resource library to all AI clusters based on scheduled tasks or data update events; A vector update unit, configured to collect the feature vector values returned by each AI cluster and update the matching weights of the feature model library based on the feature vector values; The cache computing unit uses distributed storage technology to temporarily store the updated feature vector values and directly calls the cached results in subsequent request processing to reduce computing time.

9. The system according to claim 6, wherein: The cache calculation unit of the asynchronous information synchronization module adopts distributed storage technology to achieve fast reading and writing and redundant backup of feature vector values.

10. A computer storage medium, characterized in that The computer storage medium is used to store computer execution instructions; the computer execution instructions are used to execute the intelligent fusion platform scheduling method based on artificial intelligence cluster according to any one of claims 1 to 5.

Images